[diffusion] multi-platform: add Sage Attention 3 Support for sm 120 (RTX5090) (#15382)

Co-authored-by: Mengxi Li <marcyleemx@gmail.com>

python/sglang/multimodal_gen/runtime/utils/common.py

@@ -269,6 +269,13 @@ def is_blackwell():
     return torch.cuda.get_device_capability()[0] == 10
 
 
+@lru_cache(maxsize=1)
+def is_sm120():
+    if not is_cuda():
+        return False
+    return torch.cuda.get_device_capability()[0] == 12
+
+
 @lru_cache(maxsize=1)
 def is_hpu() -> bool:
     return hasattr(torch, "hpu") and torch.hpu.is_available()